Value-based branch prediction: an alternative approach to branch prediction

One of the major performance obstacles for modern microprocessor architectures focused on increasing Instruction Level Parallelism are control dependencies. In particular, in all those processors characterized by a deep pipeline and a large issue width, the pressure is on the ability of branch predictors to fetch a large number of instructions from the correct execution path. An accurate predictor can reduce control dependencies and provide a smooth flow of execution. However, some branches are not easily predicted by traditional dynamic prediction techniques, which construct a history of the previous branch outcomes. Value-based branch prediction tries to correlate these hard-to-predict branches with any other architectural or execution information, in order to improve the prediction accuracy. The purpose of this work is to analyze some value-based branch predictors presented in the scientific literature and evaluate their accuracy and performance. The report is organized as follows: the first chapter introduces control hazards and value-based prediction through the literature. The second and the third chapters focus on the software implementation of the chosen predictor, first statically - through statistics - then dynamically - modifying the execution flow. The RTL implementation is addressed in the fourth chapter, while the fifth one contains some final remarks and propositions for further work.